{
  "reviewed_on": "2026-10-21",
  "method": "Primary research and first-party operational reports. Company metrics retain their reported periods and status. L3 classifications and economic mechanisms are the author's interpretation. Scenario inputs are illustrative and independent of company results.",
  "sources": [
    {
      "id": "support",
      "title": "Generative AI at Work",
      "url": "https://doi.org/10.1093/qje/qjae044",
      "type": "Peer-reviewed field study",
      "period_or_version": "Quarterly Journal of Economics, May 2025",
      "finding_used": "5,172 customer support agents; 15% average increase in issues resolved per hour.",
      "boundary": "Assistance within a single firm's support operation. Does not establish enterprise autonomy or general employment effects."
    },
    {
      "id": "teammate",
      "title": "The Cybernetic Teammate",
      "url": "https://www.nber.org/papers/w33641",
      "type": "Preregistered randomized field experiment",
      "period_or_version": "NBER Working Paper 33641, 2025",
      "finding_used": "776 P&G professionals; individuals with AI matched teams without AI on the studied product innovation challenges.",
      "boundary": "Task performance in a bounded experiment; not the elimination of specialist functions or a company-wide maturity classification."
    },
    {
      "id": "patterns",
      "title": "Shifting Work Patterns with Generative AI",
      "url": "https://www.microsoft.com/en-us/research/publication/shifting-work-patterns-with-generative-ai/",
      "type": "Cross-industry randomized field experiment, author research summary",
      "period_or_version": "April 2025 summary; approximately 6,000 workers over six months",
      "finding_used": "Email work changed, while time spent in meetings did not significantly change.",
      "boundary": "The cited summary identifies an early version. Later NBER revisions report a different sample; no sample or estimate is combined across versions."
    },
    {
      "id": "jcurve",
      "title": "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies",
      "url": "https://www.nber.org/papers/w25148",
      "type": "Economic model and empirical analysis of complementary intangible investment",
      "period_or_version": "2018, revised 2020",
      "finding_used": "Complementary processes, skills and other intangible investments can precede measured productivity benefits.",
      "boundary": "Does not estimate an L3 threshold or a particular company's return on AI."
    },
    {
      "id": "ibm",
      "title": "IBM AskHR",
      "url": "https://www.ibm.com/case-studies/ibm-askhr",
      "type": "Company-reported internal operating case",
      "period_or_version": "Transactions in 2024; cost reduction over four years; agentic enhancement in 2025",
      "finding_used": "More than one million HR transactions in 2024; 94% containment of common questions; contribution to a 40% reduction in HR operational costs over four years. Integrations include Workday, SAP and Concur.",
      "boundary": "Different denominators and periods. Containment is not proof of verified transaction success; the entire multi-year saving cannot be attributed to the 2025 platform enhancement."
    },
    {
      "id": "cisco",
      "title": "Cisco and Mistral AI Partner to Deliver on Cisco's Mission to Transform the Customer Experience with AI",
      "url": "https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2025/m02/cisco-and-mistral-ai-partner-to-deliver-on-cisco-mission-to-transform-the-customer-experience-with-ai.html",
      "type": "Company product and deployment announcement",
      "period_or_version": "10 February 2025",
      "finding_used": "Renewal preparation combines more than 50 signals and sources. Estimated up to 20% reduction in preparation time.",
      "boundary": "The time result is an estimate, not a realized company-wide cost saving. The announcement does not establish autonomous execution of the entire renewal."
    },
    {
      "id": "dow",
      "title": "AI impact at Dow: Copilot identifies millions in cost savings",
      "url": "https://www.microsoft.com/en-us/worklab/ai-impact-at-dow-copilot-identifies-millions-in-cost-savings",
      "type": "Vendor-published customer case with named Dow participants",
      "period_or_version": "Proof-of-concept report",
      "finding_used": "More than 100,000 annual PDF shipping invoices; automated extraction and anomaly surfacing, followed by employee investigation using a second agent.",
      "boundary": "The report anticipates savings of millions after global scaling. It does not establish that those savings were realized, or that disputes and payments were executed autonomously."
    },
    {
      "id": "klarna",
      "title": "Klarna Group plc 2025 prospectus, Form 424B4",
      "url": "https://www.sec.gov/Archives/edgar/data/2003292/000200329225000052/klarnagroupplc424b4.htm",
      "type": "Company securities filing; management operating estimates",
      "period_or_version": "69% chat share for the twelve months ended June 2025; approximately $39 million cost savings in calendar 2024",
      "finding_used": "Large-scale AI customer service alongside continued high-quality human support.",
      "boundary": "Company-reported estimates, not an independent randomized causal assessment. The periods of the two metrics differ."
    },
    {
      "id": "nist",
      "title": "AI Risk Management Framework 1.0 Core",
      "url": "https://airc.nist.gov/airmf-resources/airmf/5-sec-core/",
      "type": "Voluntary technical risk-management framework",
      "period_or_version": "AI RMF 1.0, 2023",
      "finding_used": "Independent review, deployment-relevant assessment and monitoring of production behavior.",
      "boundary": "Guidance for evaluation and governance, not proof of financial returns or an autonomy certification."
    },
    {
      "id": "alphaevolve",
      "title": "AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms",
      "url": "https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/",
      "type": "First-party research and operational report",
      "period_or_version": "14 May 2025",
      "finding_used": "Automated evaluation of evolving algorithms with discovered algorithms deployed in Google's infrastructure.",
      "boundary": "A bounded improvement mechanism with deployed outputs; not company-wide L4 or unrestricted production self-modification."
    },
    {
      "id": "harness",
      "title": "Harness²: Recursive Agent Harnessing for an Open World",
      "url": "https://arxiv.org/html/2610.14224v1",
      "type": "Research preprint with benchmark evaluation",
      "period_or_version": "arXiv:2610.14224v1, 9 October 2026",
      "finding_used": "Recursive execution-harness refinement and retention of improvement experience with frozen model parameters.",
      "boundary": "Benchmark research rather than evidence of company-wide L4. Proxy feedback and additional optimization costs remain material limitations."
    }
  ]
}
